Colonoscopy is a gold standard procedure but is highly operator-dependent. Efforts have been made to automate the detection and segmentation of polyps, a precancerous precursor, to effectively minimize missed rate. Widely used computer-aided polyp segmentation systems actuated by encoder-decoder have achieved high performance in terms of accuracy. However, polyp segmentation datasets collected from varied centers can follow different imaging protocols leading to difference in data distribution. As a result, most methods suffer from performance drop and require re-training for each specific dataset. We address this generalizability issue by proposing a global multi-scale residual fusion network (GMSRF-Net). Our proposed network maintains high-resolution representations while performing multi-scale fusion operations for all resolution scales. To further leverage scale information, we design cross multi-scale attention (CMSA) and multi-scale feature selection (MSFS) modules within the GMSRF-Net. The repeated fusion operations gated by CMSA and MSFS demonstrate improved generalizability of the network. Experiments conducted on two different polyp segmentation datasets show that our proposed GMSRF-Net outperforms the previous top-performing state-of-the-art method by 8.34% and 10.31% on unseen CVC-ClinicDB and unseen Kvasir-SEG, in terms of dice coefficient.
Recurrent neural networks (RNNs) are a popular choice for modeling sequential data. Modern RNN architectures assume constant time-intervals between observations. However, in many datasets (e.g. medical records) observation times are irregular and can carry important information. To address this challenge, we propose continuous recurrent units (CRUs) -- a neural architecture that can naturally handle irregular intervals between observations. The CRU assumes a hidden state, which evolves according to a linear stochastic differential equation and is integrated into an encoder-decoder framework. The recursive computations of the CRU can be derived using the continuous-discrete Kalman filter and are in closed form. The resulting recurrent architecture has temporal continuity between hidden states and a gating mechanism that can optimally integrate noisy observations. We derive an efficient parameterization scheme for the CRU that leads to a fast implementation f-CRU. We empirically study the CRU on a number of challenging datasets and find that it can interpolate irregular time series better than methods based on neural ordinary differential equations.
Grounded Situation Recognition (GSR) is the task that not only classifies a salient action (verb), but also predicts entities (nouns) associated with semantic roles and their locations in the given image. Inspired by the remarkable success of Transformers in vision tasks, we propose a GSR model based on a Transformer encoder-decoder architecture. The attention mechanism of our model enables accurate verb classification by capturing high-level semantic feature of an image effectively, and allows the model to flexibly deal with the complicated and image-dependent relations between entities for improved noun classification and localization. Our model is the first Transformer architecture for GSR, and achieves the state of the art in every evaluation metric on the SWiG benchmark. Our code is available at https://github.com/jhcho99/gsrtr .
The field-of-view is an important metric when designing a model for semantic segmentation. To obtain a large field-of-view, previous approaches generally choose to rapidly downsample the resolution, usually with average poolings or stride 2 convolutions. We take a different approach by using dilated convolutions with large dilation rates throughout the backbone, allowing the backbone to easily tune its field-of-view by adjusting its dilation rates, and show that it's competitive with existing approaches. To effectively use the dilated convolution, we show a simple upper bound on the dilation rate in order to not leave gaps in between the convolutional weights, and design an SE-ResNeXt inspired block structure that uses two parallel $3\times 3$ convolutions with different dilation rates to preserve the local details. Manually tuning the dilation rates for every block can be difficult, so we also introduce a differentiable neural architecture search method that uses gradient descent to optimize the dilation rates. In addition, we propose a lightweight decoder that restores local information better than common alternatives. To demonstrate the effectiveness of our approach, our model RegSeg achieves competitive results on real-time Cityscapes and CamVid datasets. Using a T4 GPU with mixed precision, RegSeg achieves 78.3 mIOU on Cityscapes test set at $37$ FPS, and 80.9 mIOU on CamVid test set at $112$ FPS, both without ImageNet pretraining.
We propose a character-based nonautoregressive GEC approach, with automatically generated character transformations. Recently, per-word classification of correction edits has proven an efficient, parallelizable alternative to current encoder-decoder GEC systems. We show that word replacement edits may be suboptimal and lead to explosion of rules for spelling, diacritization and errors in morphologically rich languages, and propose a method for generating character transformations from GEC corpus. Finally, we train character transformation models for Czech, German and Russian, reaching solid results and dramatic speedup compared to autoregressive systems. The source code is released at https://github.com/ufal/wnut2021_character_transformations_gec.
We consider the attributes of a point cloud as samples of a vector-valued volumetric function at discrete positions. To compress the attributes given the positions, we compress the parameters of the volumetric function. We model the volumetric function by tiling space into blocks, and representing the function over each block by shifts of a coordinate-based, or implicit, neural network. Inputs to the network include both spatial coordinates and a latent vector per block. We represent the latent vectors using coefficients of the region-adaptive hierarchical transform (RAHT) used in the MPEG geometry-based point cloud codec G-PCC. The coefficients, which are highly compressible, are rate-distortion optimized by back-propagation through a rate-distortion Lagrangian loss in an auto-decoder configuration. The result outperforms RAHT by 2--4 dB. This is the first work to compress volumetric functions represented by local coordinate-based neural networks. As such, we expect it to be applicable beyond point clouds, for example to compression of high-resolution neural radiance fields.
The sewerage infrastructure is one of the most important and expensive infrastructures in modern society. In order to efficiently manage the sewerage infrastructure, automated sewer inspection has to be utilized. However, while sewer defect classification has been investigated for decades, little attention has been given to classifying sewer pipe properties such as water level, pipe material, and pipe shape, which are needed to evaluate the level of sewer pipe deterioration. In this work we classify sewer pipe defects and properties concurrently and present a novel decoder-focused multi-task classification architecture Cross-Task Graph Neural Network (CT-GNN), which refines the disjointed per-task predictions using cross-task information. The CT-GNN architecture extends the traditional disjointed task-heads decoder, by utilizing a cross-task graph and unique class node embeddings. The cross-task graph can either be determined a priori based on the conditional probability between the task classes or determined dynamically using self-attention. CT-GNN can be added to any backbone and trained end-to-end at a small increase in the parameter count. We achieve state-of-the-art performance on all four classification tasks in the Sewer-ML dataset, improving defect classification and water level classification by 5.3 and 8.0 percentage points, respectively. We also outperform the single task methods as well as other multi-task classification approaches while introducing 50 times fewer parameters than previous model-focused approaches. The code and models are available at the project page http://vap.aau.dk/ctgnn
Editing and manipulating facial features in videos is an interesting and important field of research with a plethora of applications, ranging from movie post-production and visual effects to realistic avatars for video games and virtual assistants. Our method supports semantic video manipulation based on neural rendering and 3D-based facial expression modelling. We focus on interactive manipulation of the videos by altering and controlling the facial expressions, achieving promising photorealistic results. The proposed method is based on a disentangled representation and estimation of the 3D facial shape and activity, providing the user with intuitive and easy-to-use control of the facial expressions in the input video. We also introduce a user-friendly, interactive AI tool that processes human-readable semantic labels about the desired expression manipulations in specific parts of the input video and synthesizes photorealistic manipulated videos. We achieve that by mapping the emotion labels to points on the Valence-Arousal space (where Valence quantifies how positive or negative is an emotion and Arousal quantifies the power of the emotion activation), which in turn are mapped to disentangled 3D facial expressions through an especially-designed and trained expression decoder network. The paper presents detailed qualitative and quantitative experiments, which demonstrate the effectiveness of our system and the promising results it achieves.
This work proposes a new computational framework for learning a structured generative model for real-world datasets. In particular, we propose to learn a closed-loop transcription between a multi-class multi-dimensional data distribution and a linear discriminative representation (LDR) in the feature space that consists of multiple independent multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as the equilibrium point of a two-player minimax game between the encoder and decoder. A natural utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation draws inspiration from closed-loop error feedback from control systems and avoids expensive evaluating and minimizing approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, this new formulation unifies the concepts and benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark imagery datasets demonstrate tremendous potential of this new closed-loop formulation: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and often better than existing methods based on GAN, VAE, or a combination of both. Unlike existing generative models, the so learned features of the multiple classes are structured: different classes are explicitly mapped onto corresponding independent principal subspaces in the feature space. Source code can be found at https://github.com/Delay-Xili/LDR.
Complex spectrum and magnitude are considered as two major features of speech enhancement and dereverberation. Traditional approaches always treat these two features separately, ignoring their underlying relationship. In this paper, we propose Uformer, a Unet based dilated complex & real dual-path conformer network in both complex and magnitude domain for simultaneous speech enhancement and dereverberation. We exploit time attention (TA) and dilated convolution (DC) to leverage local and global contextual information and frequency attention (FA) to model dimensional information. These three sub-modules contained in the proposed dilated complex & real dual-path conformer module effectively improve the speech enhancement and dereverberation performance. Furthermore, hybrid encoder and decoder are adopted to simultaneously model the complex spectrum and magnitude and promote the information interaction between two domains. Encoder decoder attention is also applied to enhance the interaction between encoder and decoder. Our experimental results outperform all SOTA time and complex domain models objectively and subjectively. Specifically, Uformer reaches 3.6032 DNSMOS on the blind test set of Interspeech 2021 DNS Challenge, which outperforms all top-performed models. We also carry out ablation experiments to tease apart all proposed sub-modules that are most important.
This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.
Vision-and-Language Navigation (VLN) is a task that an agent is required to follow a language instruction to navigate to the goal position, which relies on the ongoing interactions with the environment during moving. Recent Transformer-based VLN methods have made great progress benefiting from the direct connections between visual observations and the language instruction via the multimodal cross-attention mechanism. However, these methods usually represent temporal context as a fixed-length vector by using an LSTM decoder or using manually designed hidden states to build a recurrent Transformer. Considering a single fixed-length vector is often insufficient to capture long-term temporal context, in this paper, we introduce Multimodal Transformer with Variable-length Memory (MTVM) for visually-grounded natural language navigation by modelling the temporal context explicitly. Specifically, MTVM enables the agent to keep track of the navigation trajectory by directly storing previous activations in a memory bank. To further boost the performance, we propose a memory-aware consistency loss to help learn a better joint representation of temporal context with random masked instructions. We evaluate MTVM on popular R2R and CVDN datasets, and our model improves Success Rate on R2R unseen validation and test set by 2% each, and reduce Goal Process by 1.6m on CVDN test set.
Real-time data collection and analysis in large experimental facilities present a great challenge across multiple domains, including high energy physics, nuclear physics, and cosmology. To address this, machine learning (ML)-based methods for real-time data compression have drawn significant attention. However, unlike natural image data, such as CIFAR and ImageNet that are relatively small-sized and continuous, scientific data often come in as three-dimensional data volumes at high rates with high sparsity (many zeros) and non-Gaussian value distribution. This makes direct application of popular ML compression methods, as well as conventional data compression methods, suboptimal. To address these obstacles, this work introduces a dual-head autoencoder to resolve sparsity and regression simultaneously, called \textit{Bicephalous Convolutional AutoEncoder} (BCAE). This method shows advantages both in compression fidelity and ratio compared to traditional data compression methods, such as MGARD, SZ, and ZFP. To achieve similar fidelity, the best performer among the traditional methods can reach only half the compression ratio of BCAE. Moreover, a thorough ablation study of the BCAE method shows that a dedicated segmentation decoder improves the reconstruction.
The existing methods to calculate the Eigenvector Centrality(EC) tend to not be robust enough for determination of EC in low time complexity or not well-scalable for large networks, hence rendering them practically unreliable/ computationally expensive. So, it is of the essence to develop a method that is scalable in low computational time. Hence, we propose a deep learning model for the identification of nodes with high Eigenvector Centrality. There have been a few previous works in identifying the high ranked nodes with supervised learning methods, but in real-world cases, the graphs are not labelled and hence deployment of supervised learning methods becomes a hazard and its usage becomes impractical. So, we devise CUL(Centrality with Unsupervised Learning) method to learn the relative EC scores in a network in an unsupervised manner. To achieve this, we develop an Encoder-Decoder based framework that maps the nodes to their respective estimated EC scores. Extensive experiments were conducted on different synthetic and real-world networks. We compared CUL against a baseline supervised method for EC estimation similar to some of the past works. It was observed that even with training on a minuscule number of training datasets, CUL delivers a relatively better accuracy score when identifying the higher ranked nodes than its supervised counterpart. We also show that CUL is much faster and has a smaller runtime than the conventional baseline method for EC computation. The code is available at https://github.com/codexhammer/CUL.
Recognition of Arabic-like scripts such as Persian and Urdu is more challenging than Latin-based scripts. This is due to the presence of a two-dimensional structure, context-dependent character shapes, spaces and overlaps, and placement of diacritics. Not much research exists for offline handwritten Urdu script which is the 10th most spoken language in the world. We present an attention based encoder-decoder model that learns to read Urdu in context. A novel localization penalty is introduced to encourage the model to attend only one location at a time when recognizing the next character. In addition, we comprehensively refine the only complete and publicly available handwritten Urdu dataset in terms of ground-truth annotations. We evaluate the model on both Urdu and Arabic datasets and show that contextual attention localization outperforms both simple attention and multi-directional LSTM models.
Segmentation of head and neck (H\&N) tumours and prediction of patient outcome are crucial for patient's disease diagnosis and treatment monitoring. Current developments of robust deep learning models are hindered by the lack of large multi-centre, multi-modal data with quality annotations. The MICCAI 2021 HEad and neCK TumOR (HECKTOR) segmentation and outcome prediction challenge creates a platform for comparing segmentation methods of the primary gross target volume on fluoro-deoxyglucose (FDG)-PET and Computed Tomography images and prediction of progression-free survival in H\&N oropharyngeal cancer.For the segmentation task, we proposed a new network based on an encoder-decoder architecture with full inter- and intra-skip connections to take advantage of low-level and high-level semantics at full scales. Additionally, we used Conditional Random Fields as a post-processing step to refine the predicted segmentation maps. We trained multiple neural networks for tumor volume segmentation, and these segmentations were ensembled achieving an average Dice Similarity Coefficient of 0.75 in cross-validation, and 0.76 on the challenge testing data set. For prediction of patient progression free survival task, we propose a Cox proportional hazard regression combining clinical, radiomic, and deep learning features. Our survival prediction model achieved a concordance index of 0.82 in cross-validation, and 0.62 on the challenge testing data set.
Salient object detection plays an important role in many downstream tasks. However, complex real-world scenes with varying scales and numbers of salient objects still pose a challenge. In this paper, we directly address the problem of detecting multiple salient objects across complex scenes. We propose a network architecture incorporating non-local feature information in both the spatial and channel spaces, capturing the long-range dependencies between separate objects. Traditional bottom-up and non-local features are combined with edge features within a feature fusion gate that progressively refines the salient object prediction in the decoder. We show that our approach accurately locates multiple salient regions even in complex scenarios. To demonstrate the efficacy of our approach to the multiple salient objects problem, we curate a new dataset containing only multiple salient objects. Our experiments demonstrate the proposed method presents state-of-the-art results on five widely used datasets without any pre-processing and post-processing. We obtain a further performance improvement against competing techniques on our multi-objects dataset. The dataset and source code are avaliable at: https://github.com/EricDengbowen/DSLRDNet.
Empirically observed time series in physics, biology, or medicine, are commonly generated by some underlying dynamical system (DS) which is the target of scientific interest. There is an increasing interest to harvest machine learning methods to reconstruct this latent DS in a data-driven, unsupervised way. In many areas of science it is common to sample time series observations from many data modalities simultaneously, e.g. electrophysiological and behavioral time series in a typical neuroscience experiment. However, current machine learning tools for reconstructing DSs usually focus on just one data modality. Here we propose a general framework for multi-modal data integration for the purpose of nonlinear DS reconstruction and the analysis of cross-modal relations. This framework is based on dynamically interpretable recurrent neural networks as general approximators of nonlinear DSs, coupled to sets of modality-specific decoder models from the class of generalized linear models. Both an expectation-maximization and a variational inference algorithm for model training are advanced and compared. We show on nonlinear DS benchmarks that our algorithms can efficiently compensate for too noisy or missing information in one data channel by exploiting other channels, and demonstrate on experimental neuroscience data how the algorithm learns to link different data domains to the underlying dynamics.
Vision-and-language (VL) pre-training has proven to be highly effective on various VL downstream tasks. While recent work has shown that fully transformer-based VL models can be more efficient than previous region-feature-based methods, their performance on downstream tasks often degrades significantly. In this paper, we present METER, a Multimodal End-to-end TransformER framework, through which we investigate how to design and pre-train a fully transformer-based VL model in an end-to-end manner. Specifically, we dissect the model designs along multiple dimensions: vision encoders (e.g., CLIP-ViT, Swin transformer), text encoders (e.g., RoBERTa, DeBERTa), multimodal fusion module (e.g., merged attention vs. co-attention), architectural design (e.g., encoder-only vs. encoder-decoder), and pre-training objectives (e.g., masked image modeling). We conduct comprehensive experiments and provide insights on how to train a performant VL transformer. METER achieves an accuracy of 77.64% on the VQAv2 test-std set using only 4M images for pre-training, surpassing the state-of-the-art region-feature-based model by 1.04%, and outperforming the previous best fully transformer-based model by 1.6%. Notably, when further scaled up, our best VQA model achieves an accuracy of 80.54%. Code and pre-trained models are released at https://github.com/zdou0830/METER.
Speech Neuroprostheses have the potential to enable communication for people with dysarthria or anarthria. Recent advances have demonstrated high-quality text decoding and speech synthesis from electrocorticographic grids placed on the cortical surface. Here, we investigate a less invasive measurement modality in three participants, namely stereotactic EEG (sEEG) that provides sparse sampling from multiple brain regions, including subcortical regions. To evaluate whether sEEG can also be used to synthesize audio from neural recordings, we employ a recurrent encoder-decoder model based on modern deep learning methods. We find that speech can indeed be reconstructed with correlations up to 0.8 from these minimally invasive recordings, despite limited amounts of training data. In particular, the architecture we employ naturally picks up on the temporal nature of the data and thereby outperforms an existing benchmark based on non-regressive convolutional neural networks.